2025 CompX Recipients

Computational Analysis of the Social Media's Role in the Effort to Overturn the 2020 Election

nyhanpic.jpg

What role did social media platforms play in the effort to overturn the 2020 election? How effective were the defenses used by platforms at countering the spread of election and voter fraud misinformation during this period? This project, which is part of the US 2020 Facebook and Instagram Election Study, will provide the first systematic answers to these questions using novel platform data from a key moment in American political history.

From Nests to Syntax: Using Deep Learning to Probe the Origins of Language

NATE DOMINY & CAT HOBAITER

The vast expressive power of human language is reliant on large repertoires of discrete meaning-bearing units, re-combined to generate near-infinite meaning. But there is a problem, the selective pressures that favored a 'language-ready brain' are unknown. Noam Chomsky and others have argued that language was an evolutionary byproduct, or exaptation, rooted in earlier selective pressures on "private conceptual abilities;" i.e., cognitive actions that integrate motor precision, non-verbal thought, and orderly spatial reasoning. It is an influential hypothesis, but linguists have little sense of when or why these skills emerged during primate evolution. To remedy this problem, we will use a bespoke deep learning model (DeepWild) to quantify the serial forelimb movements of wild chimpanzees during (i) production of intentional gestures and (ii) construction of overnight sleeping platforms, or 'nests.' These behaviors are unique to great apes, and nest-building would seem to fulfill Chomsky's precondition––a "private conceptual ability"––for the evolution of syntax. Our goal is to determine if the action and sequence of arm movements used during nest-building are homologous with those used during gestural communication. If affirmed, the finding would have far-reaching significance, putting a spotlight on arboreal nest-building as a signal moment during human evolution.

Advancing Non-Invasive Neuromodulation for Chronic Pain with Temporal Interference Stimulation

Wager & Dehghani

Chronic pain is among the most burdensome and costly medical conditions worldwide. Current treatments are ineffective for most individuals most of the time and can carry significant risks. Recent advances in neuroscience reveal that chronic pain is, in many cases, a brain disorder, maintained by maladaptive neuroplasticity in neural circuits. Chronic pain also alters brain networks that regulate mood and motivation, often leading to co-occurring  depression, anxiety, and drug dependence.

Non-invasive brain stimulation – in particular, transcranial magnetic stimulation (TMS) and transcranial direct current stimulation (tDCS) – has emerged as a promising frontier in pain treatment. However, these techniques primarily modulate superficial brain regions, limiting their efficacy in treating deep brain circuits that are critically involved in chronic pain. In this project, we study transcranial Temporal

Murphy

 Interference Stimulation (tTIS), a groundbreaking new method that enables non-invasive modulation of deep brain regions. tTIS applies high-frequency alternating currents to multiple scalp locations simultaneously, creating a stimulation envelope at a targeted frequency and brain location. This represents a paradigm shift in neuromodulation technology. Our project is a first-in-class study of tTIs effects on pain, combining computational modeling of electrical current density with experimental studies of non-invasive deep brain stimulation on human pain.

Understanding How Brain-wide Serotonin Encodes Reward 

Nautiyal

Over the past few decades, neuroscientists have probed the function of serotonin in the brain by investigating release in one brain region at a time. This has led to a disjointed and piecemeal understanding of how serotonin modulates reward processing akin to the elephant in the Hindu parable of the blind men. We propose to take a wholistic approach to understand the role of serotonin in regulating reward-related behavior. The Neukom CompX Faculty Grant will allow us to bring together two cutting edge techniques to measure serotonin release simultaneously across the whole cortex of the mouse brain. We will use biosensors which track serotonin on a subsecond timescale in the functioning mouse brain. We will build a wide-field macroscope to image the serotonin biosensor throughout the brain and characterize the spatiotemporal dynamics of serotonin release to reward.

Characterizing Differences between Human and LLM Predictions of Language 

Emily Finn

Humans evolved language on a species level and develop language on an individual level through embodied interactions in diverse sensory and social contexts. However, most large language models (LLMs) are trained using only written text, meaning their experience of language is entirely disembodied. Our past work has demonstrated that human predictions of language are more accurate than LLMs, an advantage that scales with the amount of sensory context provided to humans (from written text to audiovisual videos of speakers). In the proposed research, we will investigate how these differences between humans and LLMs are modulated by the intention of the communicator – specifically, whether the language was intended to be read (e.g., books) versus heard (e.g., conversations). 

Edge Computing for Wearable Blood Pressure Monitoring

Kofi Odame

High blood pressure is a risk factor for heart disease, stroke and diabetes, and it is responsible for 10 million deaths each year. While there are medications and lifestyle changes for treating high blood pressure, it is a difficult condition to diagnose and monitor, making effective treatment a challenge.

One promising approach for continuously monitoring high blood pressure is bioimpedance analysis of blood vessels using a smart ring. The challenge is that bioimpedance analysis requires a sophisticated machine learning algorithm that simultaneously processes multiple streams of data in real time. Conventional computational techniques cannot meet these performance requirements within the stringent resource constraints of a smart ring. To address this problem, we will develop a custom analog neural network processor for edge computing that is 1 to 2 orders of magnitude more efficient that conventional approaches, and that can perform bioimpedance analysis within the constraints of a smart ring.

Beyond blood pressure monitoring, the edge computing technology that results from this project would be useful in a wide array of biomedical and industrial settings, including: providing real time feedback during surgical procedures; monitoring crystallization to control chemical reactions in pharmaceutical sector drug development; and predictive maintenance of smart factories.

Quantitative Deep-Learning Fluorescence Surgical Guidance with a Single-photon Time-of-Flight Sensor

Bruza

A major challenge in cancer surgery is ensuring that all malignant tissue is accurately identified and removed, as positive surgical margins occur in up to 75% of cases depending on the cancer type. This can lead to cancer recurrence and the need for further treatment. Fluorescence-guided surgery improves visual contrast between cancerous and healthy tissue by labeling tumors with fluorescent molecules and visualizing them through a surgical microscope. However, despite its promise to reduce surgeon guesswork, current fluorescence imaging methods are significantly affected by signal absorption and scattering, the optical heterogeneity of tissue, and the lack of photon transport calibration—factors that contribute to variability in image interpretation and surgical decision-making. This project introduces a new imaging approach that combines advanced sensor technology with deep learning-based 3D image processing to generate more accurate, depth-resolved images during surgery, supporting better intraoperative decisions. With support from the CompX grant, we aim to integrate this novel sensor technology and advance it toward initial clinical studies.

Information Flows in Authoritarian Regimes Distant Reading Chronological Biographies of Chinese Communist Party Leaders

Yi Lu

The biographical annals, or nianpu, is one of the oldest genres in Chinese historiography, and the Chinese Communist Party (CCP) has inherited this documentary tradition. Since the late 1980s, the party has released numerous annalistic biographies of senior leaders, including Mao Zedong and Zhou Enlai. In an era where the CCP Central Archive remains largely inaccessible to scholars, these official diaries — despite being highly curated and edited — offer crucial insights into the inner workings of the party elite. They document not only the daily activities of leaders but also reference internal debates and documents that are otherwise unavailable, solidifying their status as essential texts in the field.

While scholars have leveraged these biographies to explore various facets of PRC history, my project investigates them as subjects for both historical research and computational analysis. The CompX grant will allow me to work with Dartmouth students to develop a graph schema and a training corpus that will improve automatic extraction of historical events. In doing so, this project will contribute to a more nuanced study of political communication in the world's largest bureaucracy, reconstructing how information was selected and processed over time, and enabling further research into factional politics, policy agenda-setting, and bureaucratic gkvernance in China. Methodologically, this grant will empower historians and other social scientists to leverage LLMs for improved computational text analysis as we refine the complex yet foundational task in extracting structured, event-related information from historical texts.

AI-TextbookMate: A Method to Augment Existing Textbooks with AI-Based Interactivity +

Mikhail Gronas & Sergey Bratus

This project aims to develop a method to produce AI-based interactive tutoring tools for traditional textbooks, initially focusing on modern languages  and mathematics.  Today's education  is marked by competition between traditional textbooks and learning platforms. We all know that the platforms have been gaining ground; however, traditional textbooks are still used by most students in most educational contexts around the world. This is for a good reason. Although platforms are clearly more interactive, traditional textbooks maintain an important advantage: they offer a more structured learning trajectory, based on the accumulated history of educational methodologies from the past.  We propose to bridge the gap between textbooks and digital platforms by augmenting existing textbooks with AI-driven interactivity. AI-TextbookMate will enable teachers and students to enhance widely used textbooks with AI-based interactive tasks, exercises, tests, explanations, and conversational modules.

SpoVis 2.0: Developing a Novel Open-source Method for High-throughput Microscopy Assessment of in Situ Microbial Biodiversity

Chaudhary

On Earth, 59% of all species live in soils, the majority of which are microbes. Soil microbial diversity is generally assessed using high-throughput metabarcoding of environmental DNA (eDNA), but these methods are limited in their ability to make robust predictions about how species respond to a changing planet. Many soil microbes have highly variable quantifiable morphological traits that are linked to key ecological processes (e.g. dispersal, survival), but research linking traits to in situ data is limited because the methods are slow and laborious. A critical knowledge gap exists in global biodiversity studies because microbial ecologists lack methods to rapidly process large amounts of microscope imagery to assess microbial traits in environmental samples. This CompX project will develop a novel, open-source machine learning method to automate high-throughput acquisition of trait data from microscope imagery of complex environmental samples. Faculty and staff scientists in the Environmental Sciences Department will collaborate with scientists in the Dartmouth COBRE Institute for Biomolecular Targeting to gather a large number of high-resolution microscope images from environmental samples we collected across 20 nation-wide sites. We will then train convolutional neural networks (CNNs) to classify images using the open source TFlearn deep learning library built on the TensorFlow software framework to build and train our CNN. We aim to apply this tool to analyze aerial environmental dust samples to study the dispersal of microorganisms, including human diseases and crop pathogens.

Live Brain Map Updates for Function-Preserving Tumor Resection

Xiaoyao Fan (Engineering) and Linton Evans (DHMC)

Brain tumor surgery walks a fine line: surgeons aim to remove as much of the tumor as possible while preserving neurological functions such as motor, language and vision. Preoperative MRI and diffusion tensor imaging (DTI) provide critical maps of white matter fiber tracts, which guide surgeons in avoiding damage to eloquent brain regions. However, these maps become unreliable during surgery due to brain shift cause by factors such as fluid loss and tissue resection. This project aims to develop a computational image updating framework that accounts for intraoperative brain deformation, specifically targeting the real-time localization of fiber tracts. By integrating intraoperative stereovision and ultrasound data, measured surface and subsurface displacements are used to drive a finite element biomechanical model that deforms the preoperative images—including DTI-derived fiber tracts—into an updated representation that reflects the current surgical state. With enhanced image guidance accuracy, surgeons can pursue maximal tumor resection with greater confidence, ultimately translating to improved survival and quality-of-life outcomes.

Revitalization of Endangered Languages with AI

soroush-vosoughi

Revitalizing endangered languages is crucial for preserving global cultural heritage, yet traditional documentation and learning methods face significant challenges, especially in data-scarce scenarios. Current approaches in Natural Language Processing (NLP) and Machine Learning (ML) often overlook the specific cultural and linguistic nuances inherent to endangered languages. Our research aims to address these challenges by developing an AI-driven, scalable framework using generative models and few-shot learning techniques tailored explicitly for endangered languages. Leveraging our prior work with the historically significant Nüshu language and Native American and Native Alaskan languages, we propose integrating generative models with linguistic analysis and community-driven validation. Our approach will facilitate the creation of high-quality digital resources for language preservation, documentation, and educational initiatives. This has the potential to significantly reduce resource and time constraints associated with traditional methods and promote broader engagement in language preservation efforts globally. To achieve our goals, we will first construct modular frameworks to synthesize and validate linguistic data through generative models. Next, we will develop few-shot learning models trained on enriched, high-quality corpora for efficient language documentation and learning. Finally, we will establish an interactive community-validation framework to ensure accuracy, cultural relevance, and educational utility. Successfully implementing this research will enhance our understanding of language preservation through AI, providing invaluable resources for linguists, educators, and language communities worldwide.

Unmanned Aerial System (UAS) Hyperspectral Images to Enable Alpine Zone Ecological Research

Jonathan Chipman

Cold climate ecosystems in the Arctic and on mountains are undergoing rapid environmental change in response to human alterations of the earth's climate and biogeochemical systems. In the northeastern USA and eastern Canada, an archipelago of scattered high peaks rise above the treeline and host tundra-like landscapes ("alpine zones") dominated by cold-tolerant herbaceous plants and dwarf woody shrubs. While studies on a handful of these summits have shown that the treeline is rising (Tourville et al., 2023), much less is known about ecological changes inside the alpine zone itself, such as "alpine greening" and "alpine shrubification". These peaks are also notably affected by other forms of environmental change beyond warming, such as air quality/atmospheric deposition and extraordinarily high recreational use. For this reason, they have been proposed as sensitive biomonitors of environmental change (Kimball & Weihrauch, 2000).

With this CompX grant, we will collaborate with scientists at the northeast's premier conservation NGO, the Appalachian Mountain Club, to collect field data and extensive remote sensing imagery at key alpine biodiversity sites above treeline in New Hampshire's White Mountains. The remote sensing component will include a suite of advanced multispectral and hyperspectral imaging systems and high-resolution lidar. The resulting spectral and lidar data will be used for high-resolution mapping of ecological communities and will serve as a "resolution bridge" to connect field data to coarser-scale satellite imagery, which can then be used to study changes over time in these alpine ecosystems.

Building a Community Science Tool for Inland Flooding in Vermont

Kelly & Boke

This project is informed by community-based research in the Black, White, and Ottauquechee River basins. Our research shows a need for improved technical infrastructure for knowledge transfer at different levels of response to stormwater management and flooding disasters. With community partners, we create a multidirectional process to close this knowledge transfer gap. We use mixed methods social science research and computing technologies to inform development of this new technical infrastructure. The two objectives are: 1) to build an open-source phone application for community monitoring of stormwater infrastructure and flooding risks; 2) to build a website to enhance public access to relevant data visualization.